Gaussian Mixture Language Models for Speech Recognition

1
Gaussian Mixture Language Models for Speech
Recognition

• Mohamed Afify, Olivier Siohan and Ruhi Sarikaya

2
Introduction

• Two issues for n-gram
• Generalizability adaptability
• Generalizability
• Word class / parsing
• Measure similarity in the continuous space
• Adaptability
• Larger parameter numbers for LM
• Use continuous space to reduce parameter numbers

3
Approach

• Word
• Word vector of dimensions
• New word vector of dimensions
• History concatenation of words
• History vector N-1 words, dimensions
• New History vector

M
M
N-1

M
L
4
Approach (cont.)

• Probability density for history y given the word
  w
• Probability of word w given history y
• Smoothed n-gram or smoothed clustered n-gram

or
exponents can be used to control the dynamic
ranges of n-gram and Gaussian mixture
probabilities
5
Implementation

• Word co-occurrence matrix E
• Word i follows word j
• SVD, 100 dimensions
• To create a trigram
• Two words are stacked to form a 200-d vector
• LDA MLLT
• Reduce dimensionality to 50
• GMM Training

6
Experimental results

• 5-best rescoring

7
A discriminative training framework using n-best
speech recognition transcriptions and scores for
spoken utterance classification

• Sibel Yaman, Li Deng, Dong Yu, Ye-Yi Wang, Alex
  Acero

8
Introduction

• Conventionally, a two-phase approaches is adapted
  for SUC (spoken utterance classification) task
• ASR transcription
• Semantic classification
• It has been reported that reduction in WER (word
  error rate) do not necessarily translate into CER
  (classification error rate)
• A novel discriminative training framework for
  learning the language and classification model
  is proposed

9
DT framework Using the N-best Lists

• As long as enough words are recognized to trigger
  the correct salient phrase, the correct meaning
  is assigned to the utterance
• Using ME Classifier
• Joint association score

10
DT framework Using the N-best Lists (cont.)

• The most likely to yield the correct class is
  first extracted based on joint association score
  from N-best list
• Assign remaining sentences in the N-best list
• Assignment of sentences in the N-best list to
  classes is an effective mechanism for
  discriminating the sentence in the N-best list
  that is most likely to yield the correct class
  from those that more likely to yield other wrong
  classes

11
DT framework Using the N-best Lists (cont.)

• Discriminant function loss function
• Approximation loss

12
DT framework Using the N-best Lists (cont.)

• Assignment of class

?
13
DT framework Using the N-best Lists (cont.)

• DT of LM parameters
• DT of classifier parameters

14
Experimental Results

15
Conclusions

• A new discriminative training framework for
  spoken utterance classification was proposed
• The use of N-best transcription is motivated by
  the fact the same class is often associated with
  many variants of spoken utterances